7366 ieee transactions on geoscience and remote …

7366 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 11, NOVEMBER 2014

On the Estimation of Melt Pond Fraction on theArctic Sea Ice With ENVISAT WSM Images

Marko Mäkynen, Member, IEEE, Stefan Kern, Anja Rösel, and Leif Toudal Pedersen

Abstract—The accuracy of microwave radiometer ice concen-tration (IC) retrievals in the Arctic is degraded by melt pondson sea ice during the melting season. For the development ofIC retrieval algorithms and for the quantification of their un-certainties, data sets on the area fraction of melt ponds (fmp)are needed. fmp retrieval with optical satellite data is limited byclouds. Thus, we have studied fmp retrieval with ENVISAT wideswath mode (WSM) synthetic aperture radar (SAR) images whichhave large daily coverage over the Arctic Sea ice in 2007–2012.The WSM images used here were acquired north of the FramStrait in June–August 2009. Data on fmp were available fromthe Integrated Climate Data Center’s daily Moderate ResolutionImaging Spectroradiometer (MODIS) fmp product in a 12.5-kmgrid. Relationships between SAR σ◦ and MODIS fmp were stud-ied visually by comparing daily SAR mosaics and fmp charts andby analyzing fmp and σ◦ time series and spatially and temporallycoincident fmp and σ◦ data. The correspondence between thechanges of fmp and the σ◦ statistics is too low to suggest fmp es-timation from the WSM images. In some cases, there was a 2–3-dBσ◦ increase during the ponding period. It is assumed that the vari-ation of snow and sea ice characteristics diminishes σ◦ changesdue to the melt ponding and drainage. Good correlation betweenσ◦ and fmp has only been observed for smooth landfast first-yearice in previous studies. A very interesting observation was thelarge temporal σ◦ variations during the late melting season, whichare likely linked to the atmospherically forced freezing–meltingevents. These events may also influence radiometer IC retrievals.

Index Terms—Arctic, melt ponds, radar remote sensing, sea ice.

I. INTRODUCTION

DURING the summer time, snow and sea ice melting causethe formation of the melt ponds on the Arctic sea ice [1].

These ponds reduce the surface albedo of the sea ice coversubstantially and therefore enhance summer melt. Melt en-hancement on melt-pond-covered sea ice is not only caused bythe reduced surface albedo but also by the increased penetration

Manuscript received August 22, 2013; revised December 5, 2013 andFebruary 25, 2014; accepted March 10, 2014. Date of publication April 7, 2014;date of current version May 30, 2014. This work was supported by the projectSea Ice Climate Change Initiative funded by the European Space Agency.

M. Mäkynen is with the Finnish Meteorological Institute, FI-00101 Helsinki,Finland (e-mail: [email protected]).

S. Kern is with the Integrated Climate Data Center, Center for Earth SystemResearch and Sustainability (CEN), University of Hamburg, 20144 Hamburg,Germany (e-mail: [email protected]).

A. Rösel is with the Institute of Oceanography, University of Hamburg,20146 Hamburg, Germany (e-mail: [email protected]).

L. T. Pedersen is with the Center for Ocean and Ice, Danish MeteorologicalInstitute, 2100 Kobenhavn, Denmark (e-mail: [email protected]).

Digital Object Identifier 10.1109/TGRS.2014.2311476

of solar radiation through the sea ice, as well as via enhancedlateral sea ice melt at the melt pond edges [1], [2]. Melt pondsare typically interconnected on first- and second-year sea icebut discrete on multiyear ice (MYI) or heavily deformed icedue to topography [3]. The size of melt ponds ranges over twoorders of magnitude; small melt ponds with a size below 50 m2

are found to be most common [4].The accuracy of microwave radiometer ice concentration

(IC) retrievals is reduced by the presence of wet snow and ice,by temporary thin ice layers on the top and within the near-surface snow/ice volume due to night freeze, and especially bymelt ponds on sea ice, e.g., [5]. For the development of ICretrieval algorithms and for the determination and quantifica-tion of uncertainties in the summer time, IC data sets on seaice thermodynamic state changes and the area fraction of meltponds are needed. The inclusion of melt ponds in models hasbeen shown to provide more realistic progress of the summermelt [6], [7].

The small size of melt ponds and prevailing cloudy con-ditions in summer complicate their monitoring using satellitedata, e.g., [8]. Satellite microwave radiometers and scatterom-eters have a resolution which is too coarse, e.g., [9] and [10],while the usage of optical data is hampered by cloud cover [11],[12]. Recently, a method has been developed to estimate meltpond fraction (fmp) and IC with Moderate Resolution ImagingSpectroradiometer (MODIS) surface reflectance data [8]. AMODIS fmp and open water fraction (fow) product with a12.5-km pixel size, based on the Arctic 8-day MODIS re-flectance composite (MOD09A1), is available from the Inte-grated Climate Data Center (ICDC), University of Hamburg[13]. The 8-day MODIS reflectance composite is preferred overdaily MODIS reflectance data because the method requiresclear-sky conditions which usually can only be obtained withArctic wide spatial coverage when aggregating daily MODISreflectance data of several days—in this case 8. Daily maps ofMODIS fmp can be retrieved as well but usually show a largernumber of gaps and are therefore difficult to use for monitoringfmp at a certain location and a specific date. These dailymaps are more suitable for case and intercomparison studiesas the one presented in this paper. In addition, cloud maskingof MODIS data is typically difficult over sea ice [14], [15].Therefore, an alternative way to estimate fmp will be beneficial.

The aim of our study is to investigate the possibility toobtain fmp estimates from ENVISAT wide swath mode (WSM)synthetic aperture radar (SAR) images during the melt seasonin the Arctic. An fmp estimation method is desirable, as themethod can then be applied to the large ENVISAT WSM imagearchive over the Arctic Sea ice, acquired in 2002–2011, and alsoto RADARSAT-1 and 2 ScanSAR images acquired since 1995.

0196-2892 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

MÄKYNEN et al.: ON ESTIMATION OF MELT POND FRACTION ON ARCTIC SEA ICE 7367

Fig. 1. Test area for the SAR-based melt pond fraction retrieval study. Therectangle shows the area, and gray squares are DMI weather stations. Dots arelatitude–longitude grid points for studying the temporal behavior of the MODISmelt pond fraction and ENVISAT σ◦ data. The drift track of NPEO 2009 from(triangle) June 1 to (dot) August 31, 2009 is also shown. The circular line showsthe extent of the ENVISAT WSM pole hole. The coordinate system is polarstereographic with a midlongitude of 0E and a true-scale latitude of 70N.

Data on fmp for studying relationships between SAR backscat-tering coefficient (σ◦) and fmp are available from an ICDC’sdaily MODIS melt pond product (nonpublic) with a 12.5-kmpixel size which was processed separately for our study withsimilar methods as the standard 8-day product. Our study areacovers a 950 km × 1000 km rectangle north of the Fram Strait,Greenland, and Svalbard (see Fig. 1). The area is covered byfirst-year ice (FYI) with 30–200 cm in thickness and by MYIaccording to the Danish Meteorological Institute (DMI) icecharts. Additional data sets are red-green-blue (RGB) imagesfrom MODIS Level 1 B data, numerical weather model data(ERA-Interim), coastal weather station data, and InternationalArctic Buoy Programme (IABP) buoy data [16], as well asNorth Pole Environmental Observatory (NPEO) buoy data. Alldata sets were acquired in summer 2009.

As a background for our study, we first present hereinafter thegeneral classification of the sea ice thermodynamic states [17],[18] and the review of relevant previous studies for our work,with the focus on the radar-based remote sensing of fmp. Sinceuseful products of thermodynamic state changes of sea ice (e.g.,melt and freeze onset) based on radiometer and scatterometerdata are already available, e.g., [19]–[21], and due to the lackof needed in situ data, we did not try to develop in this studySAR algorithms for the detection of melt and freeze onsets.

A. Sea Ice Thermodynamic States

In the literature, the annual cycle of sea ice is typicallypartitioned into the following thermodynamic regimes: freeze-up, winter, early melt, melt onset, and advanced melt [17],[18]. Early melt is a transition period, starting with snow

pack metamorphism and ending when moisture is continuouslypresent in the snow. At the melt onset, water in liquid phaseoccurs in the snow pack throughout the diurnal cycle, and theice surface is moist. The melt onset is further divided intopendular (melt water is held in the interstices of the snow pack)and funicular (snow grain bonds break, and gravity drainageoccurs) regimes. During advanced melt, the snow pack is firstsaturated throughout its volume and then melts rapidly, formingmelt ponds and/or flooding on the ice surface. Advanced meltis divided into periods of melt ponding and pond drainage.The ponding on the ice surface occurs as long as the brinedrainage networks, thaw holes, and cracks in the ice allowless water to disappear through gravity drainage than what iscreated by new freshwater inputs from snow melt. The gradualdecrease of fmp can be impeded by freeze–melt cycles, surfacehydraulics, predominant winds, and precipitation events onhourly to weekly time scales, e.g., [3].

B. Previous Studies

Changes from one sea ice thermodynamic regime to anotherare typically reflected in the C-band copolarization σ◦ time se-ries [17], [18], [22]. Aside from these thermodynamic changes,also the decrease and increase of snow volumetric wetness mayhave strong effects on the measured backscattering coefficient(σ◦), e.g., [22]–[24]. The attenuation of propagating radarwaves in the snow pack increases rapidly with increasing snowwetness, e.g., [23], which decreases the contribution of icesurface scattering and volume scattering from the porous MYIsnow/ice top layer to the total scattering. For example, in asurface-based C-band scatterometer data set, both bare ice andsnow-covered FYI had larger HH-polarization σ◦ signatures(up to 5 dB) during freezing (dry layered snow pack) than undermelting (wet snow) conditions [25].

Backscattering from melt ponds (in liquid state) dependsonly on their surface roughness, or wind-wave spectra, whichis a function of wind speed, upwind fetch, and pond depth, e.g.,[26]. The fetch-induced growth of surface waves depends onthe orientation, distribution, and fmp of melt ponds and also onice surface features like hummocks, which all affect the surfaceturbulence of wind stresses. Significant scale surface roughnesshas been observed for FYI melt ponds even at wind speeds of3 m/s [26]. The measured σ◦ depends also on the incidenceangle and the angle between the radar viewing direction andthe wave orientation on the melt ponds. The modeled averageC-band HH-polarized σ◦ with 100-m transects of in situ data(e.g., fmp, pond depth) over FYI showed significant contribu-tions from fetch, pond depth, and surface type variation (pondsversus hummocks) to the modeled σ◦ [26]. The formation ofan ice skim on the melt ponds has been observed to decreaseC-band copolarized σ◦ by up to 10 dB at small incidence anglesfrom σ◦ values for ponds with rough liquid surface [25].

In general, the σ◦ for melt-pond-covered sea ice has been ob-served to be highly dynamic and the primary function of fmp andsurface wind speed (ponds in liquid state), e.g., [27] and [28].

Previously, estimations of fmp and also integrated short-wave albedo have been studied for smooth landfast FYI withinthe Canadian Arctic Archipelago using RADARSAT-1 SARimages [27], [29]. Yackel and Barber [27] found a significant


positive linear relationship between σ◦ and fmp under windyconditions (around 5.3 m/s), a weaker positive relationship un-der slightly lower wind speeds (∼3.2 m/s), and no relationshipfor light wind speeds (∼1.5 m/s). The range of fmp determinedfrom aircraft video data was from 13% to 34%. The incidenceangle and surface wind speed explained 90% of the σ◦ variationduring the melt pond season. There was a strong negative rela-tionship between σ◦ and albedo under strong wind, a slightlyweaker relationship under moderate wind speeds, and a veryweak positive relationship for light wind speeds. The resultsalso suggested that σ◦ may be more closely related to thealbedo than to fmp due to the fact that albedo results from theintegration of all surface types (snow, saturated snow, and meltponds) which contribute to the measured σ◦. Hanesiak et al.[29] presented the albedo versus σ◦ relationship for the earlymelt period and summarized the results from earlier studies forother sea ice thermodynamic regimes. RADARSAT-1-derivedsurface albedo ingested into a thermodynamic sea ice modelimproved sea ice simulations to better predict the timing oflandfast FYI breakup. Scharien et al. [30] demonstrated thatthe albedo of melt-pond-covered landfast FYI can be estimatedwith better accuracy using C-band SAR-derived copolarizationratio (fco) than using only copolarized σ◦ at larger incidenceangles. It was noted that the influence of deformation features,open water areas, and deeper melt ponds of pack ice and MYIon the validity of the rco versus albedo relationship requiresfurther investigations. Using a surface-based C-band polarimet-ric scatterometer data set, Scharien et al. [25] concluded thatrco has potential for the unambiguous detection of FYI meltpond formation and fmp as it is unique from the backgroundsnow-covered or bare ice rco as long as ponds are in liquidstate. Cross-polarized σ◦ showed potential for discriminatingthe onset and duration of freeze events during the advanced meltin the marginal ice zone. The strong dependence of both co- andcross-polarized σ◦ on wind speed and pond depth was observed.

A model for fmp retrieval over large MYI floes was de-veloped by Jeffries et al. [31] for European Remote-SensingSatellite-1 (ERS-1) SAR images. In this model, fmp is derivedby assuming the σ◦ of an MYI floe to be the linear sum ofpredetermined bare ice and melt pond σ◦ values. However, asthe effect of wind roughening of melt pond surfaces was nottaken into account as in [27], we think that the validity andaccuracy of this model for different sea ice and wind conditionsare limited.

Kern et al. [32] analyzed multifrequency σ◦ data acquiredwith a helicopter-borne scatterometer in the Arctic Ocean in latesummer/early fall 2007. Results of a classification experimentwith four different surface types, i.e., old ice, nilas, open water,and melt ponds, suggested that fmp can be retrieved, but itdepended on the microwave frequency combination used inthe classification. The fmp estimates were more realistic whenfrequencies like X- and Ku-bands were combined with C-band.Old ice and melt ponds showed similar C-band HH-polarizationσ◦ signatures. Kim et al. [33] studied melt pond mapping withhelicopter-borne very high resolution (0.3 m) X-band SARdata and a coincident TerraSAR-X stripmap mode image (6-mresolution). The helicopter-borne SAR can be used to map meltponds at a level of detail comparable to aerial photography orhigh-resolution optical satellite images. Using the TerraSAR-X

image, the detection of small melt ponds (< 130 m2) was notpossible, and fmp was significantly underestimated comparedto the helicopter-borne SAR results.

In summary, the current algorithms for fmp estimation basedon SAR data are basically linear fits between σ◦ and fmp underdifferent wind speeds, and only for landfast FYI with sufficientaccuracy. Two studies with high resolution radar data [32], [33]showed that, even with high resolution (< 10 m) spaceborneSAR images, the detection of melt ponds is difficult and theirfraction is underestimated if only single frequency and singlepolarization data are used. Thus, the prospects of successfulfmp estimation from the ENVISAT WSM images with singlefrequency (C-band), HH-polarization only, and coarse resolu-tion (around 120 m) may be limited, but as a large image archiveexists and new daily fmp comparison data are available, it isworthwhile to study the possibilities.

II. DATA SETS AND PROCESSING

In the following, the data sets for studying fmp estima-tions from ENVISAT WSM images and their processing aredescribed.

A. ENVISAT WSM Images

ENVISAT WSM images with HH-polarization were ac-quired for the study area shown in Fig. 1 for June–August2009. The typical number of images per day is eight, and thetotal number of images in the data set is 705. The study areais not totally covered with the WSM images because the orbitinclination of the ENVISAT satellite limits the acquisition ofthe images up to about 87 N (denoted here as pole hole). Theswath width of the WSM images is 406 km, the image lengthis variable (minimum of 400 km), the pixel size is 75 m, andthe spatial resolution is around 113 × 123 m [34]. A WSMimage consists of five subswaths with the following incidenceangle (θ0) ranges: SS1 16.3◦–25.9◦, SS2 25.9◦–31.0◦, SS331.0◦–35.9◦, SS4 35.9◦–39.2◦, and SS5 39.2◦–42.7◦. The wholeθ0 range is then from 16.3◦ to 42.7◦.

The preprocessing of the WSM images consisted of geo-rectification, calibration (absolute σ◦), and land masking. Theimages were rectified to a polar stereographic coordinate sys-tem with a midlongitude of 0 E and a true-scale latitude of70 N. The pixel size of the rectified images is 100 m. Theequivalent number of looks (ENL) and noise equivalent σ◦(σ◦

N )in the rectified images were studied using areas of calm waterin the Barents Sea. The ENL is around 18 for the whole θ0range. Thus, the radiometric resolution is around 0.9 dB, andthe standard deviation (std) of fading is 1.0 dB. σ◦

N dependson the subswath, and it also changes as a function of θ0 insideeach subswath. For the SS1 subswath, σ◦

N varies from −21 to−19 dB; for the subswaths SS2–SS4, it is −24.5 to −23 dB;and for the SS5, it ranges from −26 to −24.5 dB. The absoluteaccuracy of σ◦ is ±0.63 dB [34].

The large θ0 variation in the WSM images causes significantchanges to the σ◦ level and contrast which complicates thesuccessful implementation of sea ice classification algorithms.Thus, we applied a θ0 compensation method to the images,following a method developed earlier for the Baltic Sea ice


[35]. For studying the θ0 dependence of σ◦ in the study area,25 windows of size 5.1 by 5.1 km were randomly selected overhigh IC regions. Next, θ0 and σ◦ differences (Δθ0 and Δσ◦ indecibel scale) were calculated from image pairs acquired withless than one day time difference. The data were divided intoweekly, half-monthly, and monthly intervals to study the effectof melt season progression in the σ◦ versus θ0 dependencein different temporal scales. The scatter plots of σ◦ versusΔθ0 indicate a linear relationship between them, and thus, alinear regression model was fitted between them. The analysisof the slope terms showed that the monthly slope terms, i.e.,−0.23 dB/1◦ in June, −0.25 dB/1◦ in July, and −0.20 dB/1◦

in August, are good approximations to follow the temporalbehavior of the σ◦ versus θ0 dependence. The coefficient ofdetermination of the linear regressions for the monthly datavaries from 0.79 to 0.94. Recently, also Zakhvatkina et al. [36]found the linear dependence between MYI σ◦ and θ0 usingWSM images and their slope term of −0.196 dB/1◦ for winterconditions is close to our results.

The incidence angle compensated WSM images (θ0 normal-ized to 30◦) were combined into a daily updated SAR imagemosaic with 500-m pixel size; see examples in Fig. 2. Themosaic has a σ◦ range from −25.5 to 0 dB with 0.1-dB steps.

B. RGB Images From MODIS Level 1 B Data

Terra MODIS daytime L1B data with reflective 500-m bandswere acquired for June–August 2009. There are typically 11images per day for our study area. MODIS data were rectifiedto the polar stereographic coordinate system that we describedearlier, keeping the original 500-m pixel size. Raw data fromreflective bands were converted to top-of-the-atmosphere sunangle corrected reflectances. Two RGB images were calculatedfrom the MODIS reflectance data: 1) bands 3-6-7 (RGB367)(this band combination is used for distinguishing betweensnow/ice and clouds) and 2) bands 2-1-3 (RGB213) (this bandcombination is suitable for the identification of flooded/wet seaice, which can be identified by the bluish color in the images).These RGB images were used to evaluate qualitatively theaccuracy of the cloud masking in the daily fmp product andfor the visual analysis of sea ice conditions.

C. MODIS Melt Pond Fraction Product

ICDC-processed daily melt pond fraction (fmp) and openwater fraction (fow) charts from a MODIS Surface Reflectancedaily L2G Global product (MOD09GA) with 500-m pixelsize. The fmp and fow retrieval is based on different spectralbehaviors of melt ponds compared to open water and snow andice [8]. fmp and fow are derived for each cloud-free 500-mpixel by applying a spectral unmixing algorithm presented in[12], which consists of a system of linear equations describ-ing observed reflectances at MODIS bands 3 (459–479 nm),1 (620–670 nm), and 2 (841–876 nm) as functions of arealfractions of three surface types, i.e., open water, melt ponds,and snow/ice, and their constant representative reflectances. Anartificial neural network has been used to derive fmp and fowfrom this set of equations [8]. The output data are gridded toa 12.5-km polar stereographic grid. The daily fmp product hasbeen validated with different data types from local observations

Fig. 2. Daily SAR mosaic from ENVISAT WSM images acquired on (a) July 4,(b) July 14, and (c) August 4, 2009. The backscattering coefficients in the mosaicare in decibel scale and scaled to the incidence angle of 30◦. The pixel size is500 m. The coordinate system is polar stereographic with a midlongitude of 0E.


(e.g., aerial photographs and ship observations) [8]. Root-mean-square errors between the fmp product and validation fmp

varied from 3.8% to 11.2%.We use here only those fmp and fow 12.5-km grid cells

which are based on at least 90% cloud-free 500-m pixels toincrease reliability on the automatic cloud screening. The prod-uct was rectified to the study area with the nearest neighboringsampling and keeping the original 12.5-km grid resolution.Examples of fmp charts are shown in Fig. 3.

The accuracy of the cloud mask in the fmp charts wasevaluated visually by comparing daily fmp charts and seriesof RGB367 images of the same dates. The comparison showsthat the fmp charts include sometimes areas of undetectedclouds and fmp for these areas are typically either over- orunderestimated when compared to neighboring cloud-free ar-eas. However, the temporal analysis of fmp in Section V-Ashows that, despite this cloud masking error, the expected meltponding and drainage periods are clearly identifiable in the fmp

time series. The cloud masking error only increases fmp datascatter in the time series.

The accuracy of fmp is additionally influenced by the lim-itations of the initial MODIS MOD09GA product. Due to thegeneral lack of relevant atmospheric data for the Arctic region,potential sources of errors are to be assumed in the atmosphericcorrection and the influence of the viewing geometry and thesolar angles. The bidirectional reflectance distribution function(BRDF) correction for most of the areas is based on modeledresults for FYI and MYI. They are used as a priori estimates ofthe BRDF [8].

Submerged lateral areas of the ice floes may appear spectrallyas melt ponds and reduce the accuracy of the melt pond product,but this effect is depending on the floe size and dependingon the occurrence of a multiplicity of ice floes. The area ofour investigations is covered mainly by thick MYI and FYI;therefore, we neglect the misclassification of submerged lateralareas of ice floes.

D. Weather Data

Weather station data were available from three DMI stations[37]; see Fig. 1. Weather observations (e.g., air temperature andwind speed) were conducted every hour or every 3 h. Weatherdata from the Kap Morris Jesup station were available after July28, 2009, and those of the other two station were available forour entire study period.

Numerical weather prediction model data were extractedfrom ERA-Interim reanalysis data. ERA-Interim has four anal-yses per day, at 00, 06, 12, and 18 UTC. The parameters usedhere were 2-m air temperature (Ta) and 10-m U and V windcomponents. The data were retrieved at a 0.5◦ grid spacing. Thespatial resolution is approximately 75 km. The data were rec-tified to the polar stereographic coordinate system of our studyarea with a 50-km grid resolution using cubic interpolation.Wind speed (Va) was calculated from the U and V components.

E. IABP and NPEO Buoy Data

The position record data of the Arctic ice buoys are availablefrom the IABP (data set C—daily buoy positions) [16]. Gridded

Fig. 3. MODIS melt pond fraction charts on (a) July 4, (b) July 14, and(c) August 4, 2009. The pixel size is 12.5 km. The land mask is shown withmelt pond fraction of 110% and missing data, which can be clouds or openwater areas, with 103%. The coordinate system is polar stereographic with amidlongitude of 0E.


12-hourly air pressure and Ta fields (data set AB) from buoydata have not yet been released for 2009. The position data areused for an overview of sea ice drift in our study area duringJune–August 2009 and for constructing fmp and σ◦ time serieswhere ice drift is tracked.

In summer 2009, the NPEO 2009 (http://psc.apl.washington.edu/northpole/index.html) drifted into our study area (seeFig. 1). NPEO included a Cold Regions Research and Engi-neering Laboratory ice mass balance (IMB) buoy 2009A [37],and its preliminary two-hourly data records are available (e.g.,snow surface position and ice thickness). NOAA/Pacific MarineEnvironmental Laboratory (PMEL) weather station data withhourly records of Ta, pressure, near-surface wind speed (Va),and station position are also available (ftp://psc.apl.washington.edu/NPEO_Data_Archive). There was also a webcam mountedon the PMEL weather station. Sea ice conditions interpretedvisually from the webcam images are described in (http://www.arctic.noaa.gov/detect/ice-npole.shtml).

III. METHODS

We study relationships between σ◦ and MODIS-derived fmp

with the following methods: 1) MODIS fmp charts, MODISRGB images, and SAR mosaic are visually compared to iden-tify the increase and decrease of fmp in the SAR mosaic; weanalyze if these changes cause a distinguishable σ◦ change andtexture, and this could give some guidelines for the analysisof the σ◦ versus fmp relationship; 2) σ◦ and fmp time seriesare extracted at the latitude–longitude grid points shown inFig. 1 and along ice buoy tracks, and the aim is to detectsignificant σ◦ and fmp (and other sea ice parameters) changes;and 3) Statistical relationships between σ◦ and fmp are studiedwith spatially and temporally (daily) coincident σ◦ and fmp

data. As supporting information for these data analyses, seaice thermodynamic state changes were estimated from ERA-Interim and weather station Ta data as in [39] and [21]. In thefollowing, procedures for the data analyses 2) and 3) and for thethermodynamic state change estimation are discussed in detail.

A. Time Series of Melt Pond Fraction andBackscattering Coefficient

The time series of σ◦ and fmp were extracted at thelatitude–longitude grid points shown in Fig. 1; latitudes are82N, 84N, and 86N, and the longitude is from −60E to +40Ewith either 20◦ (at 86N only) or 10◦ steps, if over ocean, andalong the NPEO 2009 buoy and three IABP buoy tracks (icedrift tracked). These three IABP buoys had long position timeseries within our study area.

For the fmp time series, windows with a size of 37.5×37.5 km2 (3 × 3 12.5-km pixels) were centered at the gridpoints, and the buoy locations and the mean fmp and fow werecalculated if a window was totally cloud free.σ◦ data were extracted from the WSM images (100-m pixel

size) using a 25.1-km window (251 × 251 pixels) at thefixed grid points and a 12.5-km window centered at the buoylocations. From the window data, the following parameterswere calculated: mean θ0 and mean and std of σ◦ in decibelscale. To mitigate the effect of large scale ice features (e.g.,ridges, hummocks, and leads) on the mean σ◦, a large averaging

window was used. The σ◦ time series along the buoy tracksare examples where the ice drift has been taken into account.In the other time series, we did not follow strictly the samearea of ice but analyzed data from fixed latitude–longitude gridpoints. To diminish the effect of the large θ0 variation (from16.3◦ to 42.7◦) in the σ◦ time series, they were first divided intotwo subswath categories, i.e., SS1 (16.3◦–25.9◦) and SS2–SS5(25.9◦–42.7◦), and then scaled to the average incidence anglesof 21◦ and 34◦, respectively. For melt ponds, the σ◦ versus θ0dependence is larger than for snow covered and bare ice andalso non-linear, and it is larger than that at larger incidenceangles [25]. Thus, having two σ◦ time series with differentθ0 ranges should increase the accuracy of the θ0 scaling. Thenumber of samples in the time series ranges from 42 to 240,and the average is 137. The grid points along 86N have σ◦ dataonly where > 35.5◦ due to the proximity of the pole hole whichmeans that this area can only be observed in the far θ0 range ofthe WSM images.

It is possible to estimate the random variation in the θ0-scaledσ◦ values due to the varying true σ◦ versus θ0 dependenceusing the winter part of the σ◦ time series, when there areno significant temporal changes in sea ice characteristics. Therandom variation (σ◦

rv) is characterized here by the std of thescaled σ◦ values. Parts of the time series with significant σ◦

trends in wintertime were visually excluded from the analysis.

B. Backscattering Coefficient Statistics Versus MeltPond Fraction

For studying statistical relationships between σ◦ and fmp,the σ◦ data from the WSM images and the MODIS fmp andfow data were combined on a daily basis. First, centers of the12.5-km MODIS grid cells with valid fmp data (0%–100%)were mapped to the WSM images. Next, 12.5 km by 12.5 kmσ◦ windows (125 × 125 100-m pixels) were extracted fromthe WSM images. For each SAR window, the mean θ0 wascalculated and assigned for all pixels within the window. Fromthe σ◦ windows, the following parameters were calculated:mean σ◦, std of σ◦, modal σ◦ with 0.5-dB resolution, auto-correlation at lag one, interquartile range, skewness, kurtosis,mean of the highest one-third, and some textural features fromthe gray-level co-occurrence matrix (GLCM) [40], [41]: energy,contrast, correlation, homogeneity, entropy, cluster shade, andcluster prominence. ERA Ta and Va were also extracted for thewindows.

For the mean σ◦, the effect of the θ0 variation was compen-sated by either of the following: 1) scaling all σ◦ values to θ0 of30◦ or 2) first dividing the σ◦ data into two subswath categories,SS1 and SS2–SS5, and then scaling to the average θ0’s of 21◦

and 34◦, respectively.The textural feature energy (or angular second moment) is

a measure of gray tone transitions. Correlation is a measureof the gray tone linear dependence in the image. Contrast isa measure of the intensity contrast between a pixel and itsneighbor. Homogeneity measures the closeness of the distribu-tion of elements in the GLCM to its diagonal. The two clusterparameters (shade and prominence) emulate human perceptualbehavior. GLCM was calculated as an average of GLCMs atangles of 0◦ and 90◦ and with a distance (or offset) of one pixel.


The σ◦ data were quantized from −20.75 dB to −5 dB with a0.25-dB resolution (64 gray levels). At the 100-m pixel size, theradiometric resolution is around 0.90 dB. We did not search anoptimal quantization scheme and distance(s) as in [41].

In case of the mean σ◦ versus fmp relationship, the effectof data resolution is briefly studied by using also averagingwindow sizes of 25, 50, and 100 km for both the σ◦ andfmp data.

The number of SAR images for each day varies from 2 to 11,with an average of 8. The number of the 12.5-km SAR windowsis, on the average, 4300 for each day, and the total number ofwindows is 391 407.

The MODIS fow was used to exclude SAR windows wherethe 100-m pixels typically represent either purely open water σ◦

signatures or have large open water contributions. As IC withinpack ice is close to 100% and only at the ice edge substantiallylower than 100%, we could set the fow limit to 20% and stillretain 94% of the 12.5-km SAR windows.

C. Estimation of Sea Ice Thermodynamic State Changes

We estimated melt and freeze onsets from the ERA-Interimand weather station daily mean air temperature (Tam) by usinga 2-week running median filter on the Tam time series andfinding the days when the filtered data rise or drop below the−1 ◦C threshold chosen to represent the melting temperature ofsea ice [21], [39].

IV. GENERAL SEA ICE CONDITIONS

In summer 2009, most of the study area, north of the lineStation Nord weather station—Northwest Svalbard, was cov-ered with MYI and FYI with 30–200 cm in thickness with9/10–10/10 IC according to the DMI weekly ice charts. Icefloes were classified in size categories of > 10 km, 2–10 km,and 0.5–2 km. South of the aforementioned line, the ice typesand floe sizes were comparable to the north, but IC was slightlylower. The visual analysis of the SAR mosaic showed largeopen water areas only near Svalbard and south and southeastof the Henrik Krøyer Holme weather station (see Fig. 1). Theice edge was less than 100 km north of Svalbard. MYI and FYIpack ice was typically highly deformed with a dense network ofridges [see Fig. 2(a)]. Sea ice was generally drifting in southerlydirections. The NPEO 2009 buoy drifted around 450 km duringJune 1–August 31 (see Fig. 1). In general, the ice drift waslarger in the NE corner of the study area than in the NW one.

V. RESULTS

In the following, we first analyze the changes of sea icethermodynamic state and spatial and temporal behavior of theMODIS-derived fmp in our study area in summer 2009 and thenpresent results of our studies on the relationships between σ◦

and MODIS-derived fmp.

A. Time Series of Sea Ice Thermodynamic State and MeltPond Fraction

1) Sea Ice Thermodynamic States: In the study area (seeFig. 1), the melt onset varies from June 1, 2009 in the SE

corner to June 16 in the NW corner according to the ERA-Interim median filtered Tam data [21], [39]. The average datefor the melt onset is June 7. In situ Tam from the NPEO 2009station was below 0 ◦C until July 4, and the median filteredTam indicates melt onset on July 3. Tam from the two coastalstations, Henrik Krøyer Holme and Station Nord, shows meltonset on June 13 and 11, respectively. The NPEO 2009 IMBdata show that snow pack with the initial thickness of around50 cm started to melt at the end of June, and after 20 July,the snow thickness amounts to only a few centimeters. Thus,we assume that, within pack ice, at least the first week of Junestill represented winter ice conditions. The winter part of the σ◦

data is used to estimate the random variation in the θ0-scaledσ◦ values due to the varying true σ◦ versus θ0 dependence; seeSection III-A.

The estimated freeze onset with the ERA Tam varies fromAugust 15 to September 11 and with the mode of August 30.The NPEO 2009 Tam shows the freeze onset on August 31.At the three weather stations (from north to south order) KapMorris Jesup, Station Nord, and Henrik Krøyer Holme, thefreeze onset was August 28, September 1, and September 2,respectively. In general, the late August represented the start ofthe continuous freezing conditions.

2) Melt Pond Fractions: The only in situ data on fmp inour study area are webcam images of the PMEL weatherstation of the NPEO 2009 (see Fig. 1 for the drift track).The visual analysis of the images is described in http://www.arctic.noaa.gov/detect/ice-npole.shtml. According to this anal-ysis, melt ponds started to form on July 8, but they neverbecame very widespread, with the maximum melt pond extentobserved around July 14–16. Melt ponds were freezing over byAugust 11, but small slits of open water were still visible onSeptember 8.

Fig. 4 shows fmp and fow as a function of time from ninegrid points along 84N. The general temporal fmp trend is thefollowing: fmp increased slowly from around 10% to around15% during June. Then, within the first two weeks of July, itincreased rapidly to around 40%–45% (ponding period). Themelt pond drainage period started around July 20–25, and fmp

decreased up to mid-August to around 25%. Afterward, fmp

did not decrease further. The rather large scatter in the fmp timeseries reflects the variation of fmp between the nine grid pointswhich extend over a large geographical area.

During the first half of June which represents winter andearly melt conditions, MODIS-derived fmp was not 0% butbetween 5% and 15%. One possible explanation is that the sur-face reflectance used for snow/ice in the MODIS fmp algorithm[8] does not exactly match the true sea ice reflectance in ourstudy area which has highly deformed ice. The overestimationof low fmp (below 10%) compared to validation data wastypically observed in [8]. MODIS fow was typically below 5%until July 20 and then increased to 5%–20% in August, which,according to our qualitative visual analysis of the SAR mosaic,is somewhat too large. Without in situ data, it is difficult to findreasons for the fow level in August. In late summer, the MODISfmp algorithm likely has difficulties to distinguish open waterfrom melt ponds which have melted through the ice. It isdifficult therefore to say whether the lack of further decrease offmp after mid-August is a true trend or not. A multiyear mean


Fig. 4. (a) MODIS melt pond fraction and (b) open water fraction as afunction of time extracted from nine latitude–longitude grid points along 84N(see Fig. 1).

fmp for the entire Arctic from the 8-day MODIS product showsonly a slight decreasing trend (from 30% to 25%) in August[42]. In the second half of August, atmospheric influencescause difficulties for the fmp retrieval: fmp might increase notonly due to a real increase in melt pond coverage but alsodue to further ice melting. fmp might decrease due to newice formation in leads and on melt ponds. Snow fall on meltponds already can create a few centimeters of thick slushlike icecover, causing a substantial increase in reflectivity as observedby Kern et al. [43]. ERA daily mean Ta for the 84N gridpoints was above 0 ◦C until August 9, and after that, therewere some freezing–melting cycles until August 29 after whichthe mean Ta was continuously below 0 ◦C. The data scatterin the fmp time series does not allow a detection of freezing–melting cycles.

The fmp and fow trends along the other two latitudes aresimilar. At 86N, the ponding period was slightly longer, andthe 82N time series shows more data scatter, likely due to theproximity of the ice edge. In all fmp time series, the dailyabsolute fmp change is in 88% of the data below 5%. A fewoutliers up to 12% are observed, and they may be due to thecloud masking errors in the MODIS fmp charts.

Along the NPEO 2009 buoy track (see Fig. 1), fmp and fowhave similar trends as in Fig. 4. The ponding period startedafter the first week of July. The peak fmp with 45% wasreached around July 22. Next, fmp decreased until August 8to around 20%–25%. In June and July, fow was below 5%,and in August, fow was 10%–20%. The visual analysis of theNPEO webcam imagery showed maximum melt pond extentaround July 14–16. The main difference between the two datasets is the timing of the fmp peak. However, there are no dailyMODIS fmp data for July 15–21, so the fmp peak could havebeen earlier than on July 22.

We conclude that, despite some problems in the MODIS fmp

and fow data, overestimation at low fmp values, and too largefow in late August, the data show clearly the expected meltponding and drainage periods in our study area, and thus, it canbe used for studying fmp estimation from the ENVISAT WSMimages.

B. Visual Analysis of SAR Mosaic With MODIS Melt PondFraction Charts and RGB Images

For the visual analysis of the increasing and decreasing fmp

in the SAR mosaics, we used MODIS fmp chart and MODISRGB images from July 1 when the ponding period was startinguntil the end of pond drainage (August 15). A selection ofSAR mosaics (θ0 normalized to 30◦, 500-m pixel size) andcomparable MODIS fmp charts which contain large cloud-free areas from the aforementioned time period are shown inFigs. 2 and 3.

On July 4, fmp was still low, with values in the center ofthe image ranging from 5% to 22% and an average of 15% [seeFig. 3(a)]. An area of larger fmp near the NE coast of Greenlandis likely due to cloud contamination in the fmp retrieval. TheRGB213 image (not shown) has a whitish tone in the middleof the image suggesting mainly snow-covered sea ice. In theSAR mosaic, a dense network of ice ridges is clearly visible[see Fig. 2(a)]. σ◦ has typically large spatial variation (texture),the std of σ◦ within a 12.5-km block is typically from 1.1 to3.4 dB (90% variation interval), and the average is 2.1 dB. Incomparison, the std of fading is around 1.0 dB.

On July 14, fmp is now larger, its 90% variation intervalis from 22% to 43% over the whole cloud-free area, and theaverage is 34%. The elongated area of low fmp in the middle ofFig. 3(b) appears also in the fmp chart for July 13. This featureis therefore a real fmp anomaly and not a cloud masking error.The next cloud-free data set from this area is from July 22, andthen, fmp reaches values over 40%. The RGB213 image (notshown) has mainly a bluish tone indicating melt ponded ice orice surface covered with a thin layer of liquid water. The SARmosaic on July 14 [see Fig. 2(b)] shows less spatial σ◦ variationand typically larger σ◦ level than that on July 4, e.g., in themiddle of the mosaic, ridges are not visible and the std of σ◦ istypically only 1.1–1.9 dB.fmp decreased rapidly after July 20 until mid-August. This

period does not show clearly as a monotonous change of the σ◦

level and texture in the SAR mosaic. The only typically visiblestructures are the largest ridges and leads. The spatial variationof σ◦ is mostly small, e.g., the modal std of σ◦ for the mosaic ofAugust 4 in Fig. 2(c) is only 1.2 dB. The area of the high fmp in


Fig. 5. Time series of mean and std of σ◦ (12.5-km window) and melt pondfraction (37.5-km window) at latitude–longitude grid point at 86 N, −40 E (seeFig. 1). σ◦ data are from the ENVISAT WSM subswaths SS2–SS5. Verticallines are ERA-Interim daily mean Ta based melt and freeze onsets [21], [39].

the bottom of Fig. 3(c) is due to the unmasked clouds accordingto the RGB367 images of the same day. Outside this area, fmp

ranges from 20% to 35%, with an average of 25%.In summary, during the ponding period, the texture of σ◦

decreases, and the average σ◦ level shows only a slight increase,whereas no monotonous σ◦ texture and level changes are visiblein the drainage period. This suggests that the σ◦-based deriva-tion of fmp is difficult and the accuracy may be poor.

C. Backscattering Coefficient and Melt Pond FractionTime Series

First, we studied the random variation in the θ0-scaled σ◦

values due to the varying true σ◦ versus θ0 dependence usingthe winter part of the σ◦ time series. The random variation(σ◦

rv) is characterized by the std of the scaled σ◦ values. Partsof the time series with significant σ◦ trends in wintertime werevisually excluded from the analysis. σ◦

rv varies from 0.2 to0.5 dB, and the average is 0.3 dB. The maximum σ◦

rv of 0.5 dBis used to identify significant σ◦ changes; an absolute σ◦ changebetween two successive values more than six hours apart andexceeding 0.5 dB is taken to be above the “noise” level of theθ0 scaling. Large σ◦ changes within a shorter time interval aremore likely due to the increased uncertainty of the θ0 scalingwhen two consecutive σ◦ values have a large θ0 difference.

In the following, two grid point σ◦ time series out of a total of19 are studied in a more qualitative way to find out if significantσ◦ trends and changes correspond to changes in fmp, Ta (ERA-Interim or in situ), and sea ice thermodynamic state. The firstgrid point is located at 86N, −40E (NE part of the study area),and denoted henceforth as gp1, and the second one is at 84N,0E, denoted as gp12.

Fig. 5 shows the mean and std of σ◦ with subswath SS2–SS5data and fmp from gp1. Mean σ◦ time series with both SS1and SS2–SS5 data, together with fmp for gp12, are depictedin Fig. 6. The melt and freeze onset dates from the median-filtered ERA-Interim Tam are shown in the figures. In the gp1σ◦ time series, the starting of the melting period correspondsto a gradual σ◦ decrease of 2–3 dB during the second week

Fig. 6. Time series of mean σ◦ (12.5-km window) and melt pond fraction(37.5-km window) at latitude–longitude grid point at 84 N, 0 E (see Fig. 1). σ◦

data are from the ENVISAT WSM subswaths SS1 and SS2–SS5. Vertical linesare ERA-Interim daily mean Ta based melt and freeze onsets [21], [39].

of June. The ERA Tam-derived melt onset was June 8. Theσ◦ decrease in other time series varies from 1 to 5 dB. Theσ◦ decrease from winter to melting conditions has been alsoobserved, e.g., in [44] and [45].

In June–early July after the melt onset, some of the σ◦ timeseries have large σ◦ changes. For example, in Fig. 5, gp1 σ◦

increases by around 2.5 dB during June 18–22 and then de-creases roughly the same amount up to June 26. A second largeincreasing–decreasing σ◦ event (around 3.5 dB) occurs duringJune 26 to July 4. Our first guess for the cause of these changesis atmospherically forced freezing–melting events. Freezing ofsnow and ice top layer would increase contributions of icesurface scattering from large scale roughness features (beneaththe snow pack) and volume scattering from snow basal layerand ice top volume in the total σ◦, and melting then decreasesthem. During the first event and since 8 June, ERA Ta wasbetween −1.3 ◦C and +1.2 ◦C without any clear trends, butduring the second event, Ta decreased from around 0 ◦C onJune 26 to −1.5 ◦C in the morning of July 2 and then rapidlyincreased to +1 ◦C on July 4. Thus, it is possible that, at leastthe second event is due to a Ta change. However, it is alsopossible that these σ◦ changes are due to the ice drift which maychange ice characteristics (e.g., ice type) at a fixed location. Forexample, on June 18–23, an IABP buoy (id 83725) near gp1was drifting around 27 km but only in the northeast direction.The reversible nature of the σ◦ change (increasing–decreasingcycle) suggests the freezing–melting event as a more plausibleexplanation. Detailed in situ data would be needed for furtherstudies. In June, the std of σ◦ is typically much larger than thestd of speckle (around 1.0 dB) and has large temporal variationswhich are likely linked to changes of ice deformation, surfacetype (leads, FYI, and MYI), snow thickness, and wetness.

During the fast increase of fmp in the first half of July(ponding period) at gp1 and gp1,2 there is also a gradual σ◦

(subswaths SS2–SS5) increase of roughly 2–3 dB. The σ◦

increase in the gp12 time series from the SS1 subswath is large,around 4 dB. At small θ0, the separation between σ◦ for melt


Fig. 7. Daily SAR mosaic from ENVISAT WSM images acquired on August14, 2009 under freezing conditions. The backscattering coefficients in themosaic are in decibel scale and scaled to the incidence angle of 30◦. The pixelsize is 500 m. The coordinate system is polar stereographic with a midlongitudeof 0E.

ponds with rough surface and σ◦ for snow-covered or bare iceis larger than that at high θ0 [25]. At these two grid points, theERA Va varied from 2 to 9 m/s at the times of the WSM imageacquisitions. Yackel and Barber [27] found a significant positivelinear relationship between σ◦ and fmp over landfast FYI underwindy conditions (around 5.3 m/s), but no relationship for lightwind speeds (∼1.5 m/s). In some other σ◦ time series at 86Nand 84N and in all of 82N, there is no relationship between σ◦

and increasing fmp.Soon after the peak of fmp, the drainage of the melt ponds

started, and after the first week of August, a rather constant fmp

level, around 20%–30%, was reached at gp1 and gp12. Thisdrainage period does not show as a clear monotonous increasingor decreasing σ◦ trend in our time series. This is the case alsoin all other σ◦ time series.

During the late drainage period in August, σ◦ time serieshave very large temporal variations. For example, at gp1, σ◦

increased by nearly 6 dB from August 6 to 14 and thendecreased very rapidly by an equal amount until August 17.During this time, ERA Ta first decreased from around +1 ◦C to−2.5 ◦C on August 14 and then increased to +0.5 ◦C on August17. Ta from the Kap Morris Jesup station (see Fig. 1) was above0◦C on August 4–8, and then, Ta decreased to −4 ◦C on August8 followed by an increase to around +3 ◦C on August 17.The second increasing–decreasing σ◦ event occurred betweenAugust 18 and 28 also corresponding to a freeze–melt period.As these increasing–decreasing σ◦ events were present at allgrid points, we think that they were caused by the freeze–meltevents. The effect of these events on σ◦ is also illustrated by twoSAR mosaics in Figs. 2(c) and 7. The first mosaic under meltingconditions (August 4) has a lower level of σ◦ than the secondmosaic under freezing conditions (August 14). In August, thestd of σ◦ was mostly below 1.5 dB, indicating homogeneousbackscattering conditions, and likely, surface scattering from

Fig. 8. Time series of mean σ◦ (12.5-km window) and melt pond fraction(37.5-km window) along the NPEO 2009 buoy track. σ◦ data are from theENVISAT WSM subswaths SS2–SS5. Vertical lines are NPEO in situ dailymean Ta based melt and freeze onsets [21], [39].

wet ice/snow was dominating. The small spatial variation ofσ◦ is evident in the mosaics. A possible explanation for thehigh σ◦ during freezing conditions is a strong volume scatteringfrom the coarse grained snow-ice medium, e.g., [46]. In surface-based scatterometer measurements by Scharien et al. [25],bare ice had few decibels, and snow-covered ice had up to5 dB larger σ◦ under freezing conditions than under meltingconditions. Large σ◦ changes due to atmospheric forcing havealso been observed for the Baltic Sea ice [23]. In our σ◦ timeseries, the ERA Ta based freeze onset date in late Augusttypically matches with an increasing σ◦ trend.

The σ◦ and fmp time series along the NPEO 2009 buoy trackare depicted in Fig. 8. During the period of fmp increase, alsoσ◦ increases, around 3–4 dB. The melt pond drainage period inlate July shows a 2-dB drop in σ◦, but as drainage continuesin early August, the σ◦ increases by 7 dB until mid-August. InAugust, the NPEO buoy was close to gp12 which has a similarincreasing σ◦ trend (see Fig. 6). The start of the σ◦ raise isaround ten days later when the snow pack has melted to a fewcentimeters in thickness at the NPEO site. During August 1–15,in situ NPEO Ta shows only slight freezing periods on August2–5 (minTa − 1.5 ◦C) and Aug 9–11 (minTa − 1 ◦C). It isdifficult to find a plausible explanation for this increasing σ◦

trend. Maybe it is due to an increase of volume scattering asice top surface is gradually drying up and slowly increasing invertical extent as melt progresses. Another mechanism couldbe the surface scattering at melt pond edges which, as melt pro-gresses, can develop a freeboard of several centimeters height.In situ wind speed was below 2 m/s, suggesting a smooth watersurface and, thus, small scattering contribution from the meltponds themselves. In the σ◦ time series for the three selectedIABP buoys (one near gp1 and two close to the NE coast ofGreenland), only at one (near gp1) did the fmp increase in Julyresult in an increasing σ◦ trend. Thus, it seems that the trackingof ice drift does not always increase the correlation between σ◦

and fmp.Finally, we studied visually scatterplots (not shown) between

significant σ◦ changes (|Δσ◦| > 0.5 dB) and contemporaryfmp, Ta, and Va changes. These plots did not show any cor-relation between the changes of the variables.


In summary, the detection of melt onset seems to be possiblein the σ◦ time series based on the decreasing σ◦ trend from thewinter σ◦ values. This was also reported in, e.g., [44] and [45].The relationship between σ◦ and the increase of fmp (pondingperiod) was typically not very clear; in some cases, a 2–3-dB σ◦

increase was present, but the trend was noisy with fast temporalvariations mainly due to the “noise” of the θ0 scaling and likelyalso from the varying roughness of the melt ponds. The relation-ship between σ◦ and fmp was, in some cases, slightly better atsmaller θ0 values. The pond drainage period showed a σ◦ trend(decreasing) only in one case (NPEO buoy). The tracking of icedrift (buoy σ◦ time series) did not always improve the σ◦ versusfmp relationship. Likely, in our study area ice deformationfeatures and variation of ice types, upper ice volume charac-teristics, and snow thickness diminish σ◦ changes induced bymelt ponding and drainage. In August when little snow remainsand the MODIS fmp shows a level of 20%–30% without anysignificant trends, there are large slow σ◦ variations likelylinked to the atmospherically forced freezing–melting events.These events should also show up in brightness temperaturesignatures and, thus, possibly in the IC retrievals, which justifiestheir further study if detailed in situ data are available.

D. Backscattering Coefficient Statistics Versus MeltPond Fraction

Here, we included data only from July 1 and August 5which encompasses melt ponding and drainage periods. InJune, MODIS-derived fmp increases from about 10% to 15%and is likely overestimated as discussed in Section V-A. AfterAugust 5, there were large temporal σ◦ variations but smallfmp variations. The number of 12.5-km SAR windows is now192 643.

First, we studied the relationship between fmp and mean σ◦

for the 12.5-km windows. The scatterplot between fmp andmean σ◦ values from all WSM swaths does not show anysystematic relation between them. The correlation coefficient(r) is only 0.10 between the mean σ◦ and fmp. The 90%interval of the ERA Va in the data ranges from 1.3 to 8.6 m/s,and modal Va is 3.5 m/s. Fig. 9 shows the scatterplot betweenfmp and the SS1 subswath mean σ◦ values under three differentERA wind speed ranges: 0–3 m/s (calm light air), 4–7 m/s(light breeze), and larger than 8 m/s. Even though the separationbetween the melt pond and snow-covered or bare ice σ◦ valuesshould be larger at the smaller θ0’s [25] (SS1 subswath), rbetween fmp and mean σ◦ is not much larger; it is, at best, 0.30under the 4–7-m/s wind speed range.

The effect of resolution on the relationship between fmp andthe mean σ◦ was briefly studied using also averaging windowsof 25, 50, and 100 km for the both data sets. This further spatialaveraging did not increase the correlation between the fmp andmean σ◦, regardless of the wind speed range.

Next, the mean σ◦ probability density functions (pdfs) werecalculated for 5% wide fmp intervals (5%–9%, 10%–14%,etc.) using the data from the 12.5-km size windows. Fig. 10shows pdfs with mean σ◦ values from all WSM subswaths. Thenumber of samples for each pdf is, at minimum, 443, and theaverage is 17 475. The σ◦ pdfs for different fmp intervals forfmp larger than 25% do not differ much from each other. The

Fig. 9. Scatterplot between the melt pond fraction and the mean σ◦ at12.5-km pixel size under three different ERA-Interim wind speed ranges. Allσ◦ data are from the WSM subswath SS1 and the incidence angle scaled to21◦. Only data from July 1 to August 5, 2009 are used. MODIS open waterfraction is below 20%.

Fig. 10. PDFs of the mean σ◦ values at 12.5-km pixel size for different meltpond fraction intervals. Only data from July 1 to August 5, 2009 are used.MODIS open water fraction below 20%. The bin width of pdfs is 0.5 dB.

pdfs for fmp intervals of 10%–14%, 15%–19%, and 20%–24%have a somewhat larger proportion of smaller σ◦ values thanthe pdfs for larger fmp. Only the pdf for the fmp interval of5%–9% clearly differs from the rest; its peak position is on a2–3-dB larger σ◦ value than the peaks of other pdfs. All σ◦ datafor this pdf were acquired on July 1–4, when the areas withfmp between 5%–9% were north of Greenland and had ratherhigh average σ◦ due to freezing conditions; see σ◦ time seriesin Fig. 5. The separation of pdfs does not improve by restrictingthe θ0 range (SS1 or SS2–SS5). Mean σ◦ values from all WSMsubswaths range only from −17 to −10 dB, regardless of thefmp interval.

In summary, using σ◦ data block averaged to the same pixelsize as the MODIS fmp data (12.5 km), it is not possibleto estimate fmp. Very likely, at this coarse spatial resolution,


other properties of sea ice, mainly ice deformation featuresand mixtures of various ice types (e.g., MYI floes and brashice), are masking σ◦ variations by changing fmp, as was alsosuggested in the σ◦ time series study. For better studying therelationship between σ◦ level and fmp, we need fmp dataat higher resolution. Yackel and Barber [27] used their studyRADARSAT-1 SAR data averaged to a 500-m pixel size andfine-resolution (∼16.5m2) aircraft video-based fmp data andfound good correlation between σ◦ and fmp for smooth landfastFYI under windy conditions (around 5.3 m/s).

Finally, we studied the relationships between fmp and vari-ous statistical (e.g., std, mode, and autocorrelation) and GLCMtexture features (see Section III-B) calculated from the 12.5-kmSAR σ◦ windows with 100-m pixel size. None of the parame-ters has a significant relationship with fmp, and their pdfs donot differ significantly for different fmp intervals, regardlessof the incidence angle range selection (all subswaths, SS1, orSS2–SS5).

VI. CONCLUSION

We have studied fmp estimation over the Arctic MYI andFYI using ENVISAT WSM images acquired over an area northof the Fram Strait, Greenland, and Svalbard. Melt pond andopen water fraction data were available from ICDC’s dailyMODIS melt pond product with a 12.5-km pixel size [8]. Datasets were acquired for June–August 2009.

The MODIS fmp data showed expected melt ponding anddrainage periods in our study area. fmp increased from about10% to 15% in June. Within the first two weeks of July, itincreased rapidly to around 40%–45% (ponding period) andthen decreased up to mid-August to around 25% (drainageperiod). Afterward, fmp did not decrease further. The nonzerofmp in the beginning of June may be due to a slight mismatchof surface reflectance for snow/ice used in the MODIS fmp

algorithm for our study area with highly deformed ice. Theoverestimation of low fmp (below 10%) compared to validationdata was typically observed by Rösel et al. [8]. It is difficultto say whether the stagnation of fmp after mid-August is areal trend or not. A multiyear mean fmp for the whole Arcticfrom the 8-day MODIS product showed only a slight decreas-ing trend (from 30% to 25%) in August [42]. We note thatmelting and freezing coexist at that time of the year and thatother effects than actual changes in absolute melt pond areacan contribute to the temporal development of fmp. ERA Ta

suggests freezing–melting cycles in August which could haveslowed the melt pond drainage.

Possible relationships between SAR σ◦ and MODIS fmp

were studied visually by comparing daily SAR mosaics andfmp charts and MODIS RGB images and by analyzing fmp

and σ◦ time series and spatially and temporally coincident fmp

and σ◦ data from 12.5-km windows. The θ0 range for theσ◦ data was either that for the SS1 subswath (16.3◦–25.9◦),SS2–SS5 swaths (25.9◦–42.7◦), or all WSM swaths (16.3◦–42.7◦). Overall, the results showed only little correspondencebetween the increase and decrease of fmp and the σ◦ statistics(mean, std, and other texture parameters), Therefore, the fmp

estimation from the WSM images is not possible, at least inour data sets. In some cases, there was a 2–3-dB σ◦ increase

during the ponding period, but the trend had fast temporalvariations due to the “noise” of the θ0 scaling in σ◦ and likelyalso from the varying roughness of the melt pond surface dueto changing wind conditions. The relationship between σ◦

and fmp was, in some cases, slightly better at smaller θ0. Thetracking of ice drift with the ice buoy data did not significantlyimprove the σ◦ versus fmp relationship. Likely, in our studyarea, ice deformation features, variation of ice types, upperice volume characteristics, and snow pack diminish both σ◦

level changes at 12.5-km resolution and textural σ◦ changes at100-m resolution due to the melt ponding and drainage.

Previously, good correlation between σ◦ and fmp has onlybeen observed for smooth landfast FYI [27]. Studies with highresolution radar data [32], [33] showed that, even with highresolution (< 10 m), spaceborne SAR image detection of meltponds is difficult and fmp is underestimated if only singlefrequency and single polarization data are used. Nevertheless,it may be worthwhile to repeat our study if high-resolutionfmp data sets, at least with resolution comparable to the typ-ical 50–100-m resolution of wide swath SAR images (e.g.,RADARSAT-2 ScanSAR), with large spatial and temporal cov-erage become available and, better yet, if these are accompaniedwith high-resolution sea ice albedo data. Statistically, σ◦ maybe more closely related to the albedo than to fmp as albedoresults from the integration of all surface types (snow, bare ice,and melt ponds) which contribute to the measured σ◦[27]. Ifan algorithm for the fmp or albedo estimation over the Arcticdrift ice can be developed then fmp or albedo products couldbe calculated from a large ENVISAT WSM and RADARSAT-1/2 ScanSAR image archive over the Arctic Sea ice. Thelarge θ0 variation in the WSM images is a problem in the σ◦

analysis and image classification, as it can be compensated foronly partially. Because melt ponds are mostly smaller than theWSM image resolution of 100 m, leading to mixed pixel σ◦

signatures, a future study should also include fine-resolution(< 10−20 m) SAR images with limited θ0 ranges, preferablywith two copolarizations (HH and VV). According to surface-based C-band scatterometer measurements, the copolarizationratio has potential for unambiguous detection of FYI melt pondformation and fmp [25].

A very interesting observation was the large typically long-term temporal σ◦ variation during the late melting season inAugust when little snow remains and the MODIS fmp wasbased at a level of 20%–30% without any systematic variations.These σ◦ variations are very likely linked to the atmosphericallyforced freeze–melt events. A plausible explanation for highσ◦ in freezing conditions is strong volume scattering from thecoarse grained snow-ice medium. These events should alsoshow in brightness temperature signatures and, thus, possibly inthe radiometer IC retrievals, which justifies their further study,if detailed in situ data become available.

REFERENCES

[1] D. K. Perovich, S. V. Nghiem, T. Markus, and A. Schweiger, “Seasonalevolution and interannual variability of the local solar energy absorbedby the arctic sea ice—Ocean system,” J. Geophys. Res., vol. 112, no. C3,pp. C03005-1–C03005-13, Mar. 2007.

[2] M. Nicolaus, C. Katlein, J. Maslanik, and S. Hendricks, “Changes in arcticsea ice result in increasing light transmittance and absorption,” Geophys.Res. Lett., vol. 39, no. 24, pp. L24501-1–L24501-6, Dec. 2012.


[3] H. Eicken, T. C. Grenfell, D. K. Perovich, J. A. Richter-Menge, andK. Frey, “Hydraulic controls of summer arctic pack ice albedo,”J. Geophys. Res., vol. 109, no. C8, pp. C08007-1–C08007-12, Aug. 2004.

[4] D. K. Perovich, W. B. Tucker, III, and K. A. Ligett, “Aerial observationsof the evolution of ice surface conditions during summer,” J. Geophys.Res., vol. 107, no. C10, pp. SHE 24-1–SHE 24-14, Oct. 2002.

[5] D. J. Cavalieri, B. A. Burns, and R. G. Onstott, “Investigation of the effectsof summer melt on the calculation of sea ice concentration using activeand passive microwave data,” J. Geophys. Res., vol. 95, no. C4, pp. 5359–5369, Apr. 1990.

[6] C. A. Pedersen, E. Roeckner, M. Luethje, and J.-G. Winther, “A new seaice albedo scheme including melt ponds for ECHAM5 general circulationmodel,” J. Geophys. Res., vol. 114, no. D8, pp. D08101-1–536D08101-15, Apr. 2009.

[7] D. Flocco, D. Schroeder, D. L. Feltham, and E. C. Hunke, “Impact of meltponds on Arctic sea ice simulations from 1990 to 2007,” J. Geophys. Res.,vol. 117, no. C9, pp. C09032-1–C09032-17, Sep. 2012.

[8] A. Rösel, L. Kaleschke, and G. Birnbaum, “Melt ponds on arctic sea icedetermined from MODIS satellite data using an artificial neural network,”Cryosphere, vol. 6, no. 2, pp. 431–446, Apr. 2012.

[9] J. C. Comiso and R. Kwok, “Surface and radiative characteristics of thesummer arctic sea ice cover from multisensory satellite observation,”J. Geophys. Res., vol. 101, no. C12, pp. 28 397–28 416, Dec. 1996.

[10] F. Fetterer and N. Untersteiner, “Observations of melt ponds on arc-tic sea ice,” J. Geophys. Res., vol. 103, no. C11, pp. 24 821–24 835,Oct. 1998.

[11] T. Markus, D. J. Cavalieri, M. A. Tschudi, and A. Ivanoff, “Comparisonof aerial video and Landsat 7 data over ponded sea ice,” Remote Sens.Environ., vol. 86, no. 4, pp. 458–469, Aug. 2003.

[12] M. A. Tschudi, J. A. Maslanik, and D. K. Perovich, “Derivation of meltpond coverage on arctic sea ice using Modis observation,” Remote Sens.Environ., vol. 112, no. 5, pp. 2605–2614, May 2008.

[13] A. Rösel, L. Kaleschke, and S. Kern, “Gridded melt pond cover fraction onarctic sea ice derived from TERRA-MODIS 8-day composite reflectancedata,” Hamburg, Germany, 2013.

[14] R. A. Frey, S. A. Ackerman, Y. Liu, K. I. Strabala, H. Zhang, J. R. Key,and X. Wang, “Cloud detection with MODIS. Part I: Improvements in theMODIS cloud mask for collection 5,” J. Atmos. Ocean. Technol., vol. 25,no. 7, pp. 1057–1072, Jul. 2008.

[15] M. A. Chan and J. C. Comiso, “Arctic cloud characteristics as derivedfrom MODIS, CALIPSO, and CloudSat,” J. Clim., vol. 26, no. 10,pp. 3285–3306, May 2013.

[16] I. G. Rigor, “IABP drifting buoy pressure, temperature, position, andinterpolated ice velocity,” National Snow and Ice Data Center, Boulder,CO, USA, 2002, [Jun.–Aug. 2009].

[17] C. E. Livingstone, R. G. Onstott, L. D. Arsenault, A. L. Gray, andK. P. Singh, “Microwave sea-ice signatures near the onset of melt,”IEEE Trans. Geosci. Remote Sens., vol. GE-25, no. 2, pp. 174–187,Mar. 1987.

[18] D. G. Barber, J. J. Yackel, and J. M. Hanesiak, “Sea ice, RADARSAT-1and arctic climate processes: A review and update,” Can. J. Remote Sens.,vol. 27, no. 1, pp. 51–61, 2001.

[19] T. Markus, J. C. Stroeve, and J. Miller, “Recent changes in arctic sea icemelt onset, freezeup, and melt season length,” J. Geophys. Res., vol. 114,no. C12, pp. C12024-1–C12024-14, Dec. 2009.

[20] L. Wang, G. J. Wolken, M. J. Sharp, S. E. L. Howell, C. Derksen,R. D. Brown, T. Markus, and J. Cole, “Integrated pan—Arctic melt on-set detection from satellite active and passive microwave measurements,2000–2009,” J. Geophys. Res., vol. 116, no. D22, pp. D22103-1–D22103-15, Nov. 2011.

[21] J. Mortin, T. M. Schrøder, A. Walløe Hansen, B. Holt, andK. C. McDonald, “Mapping of seasonal freeze-thaw transitions across thepan-arctic land and sea ice domains with satellite radar,” J. Geophys. Res.,vol. 117, no. C8, pp. C08004-1–C08004-19, Aug. 2012.

[22] J. J. Yackel, D. G. Barber, T. N. Papakyriakou, and C. Breneman, “First-year sea ice spring melt transitions in the Canadian Arctic Archipelagofrom time-series synthetic aperture radar data, 1992–2002,” Hydrol.Process., vol. 21, no. 2, pp. 253–265, Jan. 2007.

[23] M. Mäkynen, B. Cheng, M. Similä, T. Vihma, and M. Hallikainen, “Com-parisons between SAR backscattering coefficient and results of a thermo-dynamic snow/ice model for the Baltic Sea land-fast sea ice,” IEEE Trans.Geosci. Remote Sens., vol. 45, no. 5, pp. 1131–1141, May 2007.

[24] M. R. Drinkwater, “LIMEX’87 ice surface characteristics: Implicationsfor C-band SAR backscatter signatures,” IEEE Trans. Geosci. RemoteSens., vol. 27, no. 5, pp. 501–513, Sep. 1989.

[25] R. K. Scharien, J. J. Yackel, D. G. Barber, M. Asplin, M. Gupta, andD. Isleifson, “Geophysical controls on C band polarimetric backscatter

from melt pond covered arctic first-year sea ice: Assessment using high-resolution scatterometry,” J. Geophys. Res., vol. 117, no. C9, pp. C00G18-1–C00G18-15, Sep. 2012.

[26] R. K. Scharien and J. J. Yackel, “Analysis of surface roughness andmorphology of first-year sea ice melt ponds: Implications for microwavescattering,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 12, pp. 2927–2939, Dec. 2005.

[27] J. J. Yackel and D. G. Barber, “Melt ponds on sea ice in the CanadianArctic Archipelago. Part 2. On the use of RADARSAT-1 synthetic aper-ture radar for geophysical inversion,” J. Geophys. Res., vol. 105, no. C9,pp. 22 061–22 070, Sep. 2000.

[28] D. G. Barber and J. J. Yackel, “The physical, radiative and microwavescattering characteristics of melt ponds on arctic landfast sea ice,” Int. J.Remote Sens., vol. 20, no. 10, pp. 2069–2090, Jan. 1999.

[29] J. Hanesiak, J. Yackel, and D. Barber, “Effect of melt ponds on first-year sea ice ablation—Integration of RADARSAT-1 and thermodynamicmodeling,” Can. J. Remote Sens., vol. 27, no. 5, pp. 433–442, 2001.

[30] R. K. Scharien, J. J. Yackel, M. A. Granskog, and B. G. T. Else, “Coinci-dent high resolution optical-SAR image analysis for surface albedo esti-mation of first-year sea ice during summer melt,” Remote Sens. Environ.,vol. 111, no. 2/3, pp. 160–171, Nov. 2007.

[31] M. O. Jeffries, K. Schwartz, and S. Li, “Arctic ocean melt season charac-teristics and sea ice melt pond fractions using ERS-1 SAR,” in Proc. 4thSymp. Remote Sens. Polar Environ., 1996, pp. 127–132, ESA SP-391.

[32] S. Kern, M. Brath, and D. Stammer, “Melt ponds as observed with ahelicopter-borne, multi-frequency scatterometer in the Arctic Ocean in2007,” presented at the European Space Agency Living Planet Symp.,Bergen, Norway, Jun. 29–Jul. 2, 2010, ESA-SP-686.

[33] D.-J. Kim, B. Hwang, K. H. Chung, S. H. Lee, H.-S. Jung, andW. M. Moon, “Melt pond mapping with high-resolution SAR: The firstview,” Proc. IEEE, vol. 101, no. 3, pp. 748–758, Mar. 2013.

[34] “ENVISAT ASAR monthly report,” Paris, France, ENVI-CLVL-EOPG-TN-04-0009, Jun. 2009, Technical note.

[35] M. Mäkynen, A. T. Manninen, M. Similä, J. Karvonen, andM. Hallikainen, “Incidence angle dependence of the statistical propertiesof C-band HH-polarization backscattering signatures of the Baltic Seaice,” IEEE Trans. Geosci. Remote Sens., vol. 40, no. 12, pp. 2593–2605,Dec. 2002.

[36] N. Y. Zakhvatkina, V. Y. Alexandrov, O. M. Johannessen, and S. Sandven,“Classification of sea ice types in ENVISAT synthetic aperture radarimages,” IEEE Trans. Geosci. Remote Sens., vol. 51, no. 5, pp. 2587–2600, May 2013.

[37] “Weather and climate data from Greenland 1958–2011—Observation datawith description,” Danish Meteorol. Inst., Copenhagen, Denmark, DMITech. Rep. 12-15, 2012, J. Cappalen (editor).

[38] D. K. Perovich, J. A. Richter-Menge, B. Elder, K. Claffey, andC. Polashenski, Observing and Understanding Climate Change: Monitor-ing the Mass Balance, Motion, and Thickness of Arctic Sea Ice, 2009.[Online]. Available: http://imb.crrel.usace.army.mil/

[39] I. G. Rigor, R. L. Colony, and S. Martin, “Variations in surface air tem-perature observations in the Arctic, 1979–97,” J. Clim., vol. 13, no. 5,pp. 896–914, Mar. 2000.

[40] J. A. Nystuen and F. W. Garcia, Jr., “Sea ice classification using SARbackscatter statistics,” IEEE Trans. GeoSci. Remote Sens., vol. 30, no. 3,pp. 502–509, May 1992.

[41] L. K. Soh and C. Tsatsoulis, “Texture analysis of SAR sea ice imageryusing gray level co-occurrence matrices,” IEEE Trans. Geosci. RemoteSens., vol. 37, no. 2, pp. 780–795, Mar. 1999.

[42] A. Rösel and L. Kaleschke, “Exceptional melt pond occurrence in theyears 2007 and 2011 on the arctic sea ice revealed from MODIS satel-lite data,” J. Geophys. Res., vol. 117, no. C5, pp. C05018-1–C05018-8,May 2012.

[43] S. Kern, G. Spreen, and A. Winderlich, “Sea ice radar backscatter mea-surements for improved melt pond and thin-ice cover analysis,” in TheExpedition ARKTIS-XXII/2 of the Research Vessel “Polarstern”, vol. 579,Reports on Polar and Marine Research, U. Schauer, Ed. Bremerhaven,Germany: Alfred Wegener Institute for Polar and Marine Research,2008, 271.

[44] D. P. Winebrenner, E. D. Nelson, R. Colony, and R. D. West, “Observa-tions of melt onset on multi-year arctic sea ice using the ERS-1 SAR,”J. Geophys. Res., vol. 99, no. C11, pp. 22 425–22 442, Nov. 1994.

[45] R. Kwok, G. F. Cunningham, and S. V. Nghiem, “A study of the onset ofmelt over the arctic ocean in RADARSAT synthetic aperture radar data,”J. Geophys. Res., vol. 108, no. C11, pp. 27-1–27-13, Nov. 2003.

[46] A. Carlström and L. M. H. Ulander, “C-Band backscatter signatures ofold sea ice in the central arctic during freeze-up,” IEEE Trans. Geosci.Remote Sens., vol. 31, no. 4, pp. 819–829, Jul. 1993.


Marko Mäkynen (S’98–M’01) received the Dr.Sci.(Technology) degree from the Helsinki University ofTechnology (TKK), Espoo, Finland, in 2007.

Since 2009, he has been a Senior Scientist at theMarine Research Unit of the Finnish MeteorologicalInstitute, Helsinki, Finland. He was a SeniorTeaching Assistant at the Laboratory of SpaceTechnology, TKK, in 1999–2006 and was the ActingDirector of the laboratory in 2007–2008. His mainresearch interest is microwave and optical remotesensing of the Baltic and Arctic Sea ice.

Stefan Kern received the Diploma degree in meteo-rology from the University of Hannover, Hannover,Germany, in 1997 and the Ph.D. degree in physicsfrom the University of Bremen, Bremen, Germany,in 2001.

Since 2001, he has been with the University ofHamburg, Hamburg, Germany, where he first workedin the Institute of Oceanography and where hejoined the Center of Excellence for Climate Sys-tem Analysis and Prediction in 2010, which nowis part of the Zentrum für Erdsystemforschung und

Nachhaltigkeit. His main research interests are remote sensing of the marinecryosphere, algorithm development, and uncertainty assessment for improvedquantification of ocean–sea-ice–atmosphere interactions. He carried out fieldexperiments measuring multifrequency radar backscatter of Arctic sea ice andAlpine snow using a helicopter-borne scatterometer. He currently dedicates hiswork to sea-ice-related climate data record uncertainty assessment.

Anja Rösel was born in in Munich, Germany, onApril 22, 1976. She received the Dr.Nat.Sci. degreefrom the Institute of Oceanography, University ofHamburg, Hamburg, Germany, in 2012.

Since 2013, she has obtained a fellowship ofthe Young Researcher’s Initiative of the Universityof Hamburg. She studied physical geography withmain focus on climatology and remote sensing atthe University of Munich, Munich. From 2005 to2007, she was a member of the wintering team of theGerman Neumayer Research Station in Antarctica

and was responsible for the Meteorological Observatory. After the wintering,she started her Ph.D. at the University of Hamburg, investigating in thedetection of melt ponds from optical satellite data. Her main research interestis remote sensing of the polar regions.

Leif Toudal Pedersen was born in Denmark in 1957.He received the M.S. degree in microwave engi-neering and the Ph.D. degree in passive microwaveremote sensing of sea ice from the Technical Univer-sity of Denmark (TUD), Kongens Lyngby, Denmark,in 1982 and 1992, respectively.

From 1982 to 2000, he was a Research Assistantat the Electromagnetics Institute at TUD, and from2000 to 2007, he was an Associate Professor atØrsted-TUD, Denmark. Since 2007, he has been aSenior Researcher at the Center for Ocean and Ice at

the Danish Meteorological Institute. His research interests include the retrievalof ice, ocean, and atmospheric parameters from multispectral microwave ra-diometer measurements as well as other methods for remote sensing of sea ice.

Dr. Pedersen is a member of the Danish National Committee for ClimateResearch and the Danish National Committee for Scientific Committee onAntarctic Research. He also serves as a Danish delegate to European SpaceAgency’s Programme Board for Earth Observations and on ESA’s Sentinel-1Mission Advisory Group.

7366 ieee transactions on geoscience and remote …

Documents